sgd_op.cu 7.4 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
L
liaogang 已提交
2

L
Luo Tao 已提交
3 4 5
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
L
liaogang 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
L
liaogang 已提交
8

L
Luo Tao 已提交
9 10 11 12 13
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
L
liaogang 已提交
14

C
chengduo 已提交
15
#include <algorithm>
16
#include "paddle/fluid/operators/amp/fp16_type_traits.h"
W
Wu Yi 已提交
17
#include "paddle/fluid/operators/optimizers/sgd_op.h"
18
#include "paddle/fluid/platform/device/gpu/gpu_primitives.h"
Q
qijun 已提交
19 20 21 22 23

namespace paddle {
namespace operators {

namespace {
C
chengduoZH 已提交
24

25 26 27 28 29
template <typename T, typename MT>
__global__ void SGDKernelMT(const T* param, const T* grad,
                            const T* learning_rate, const int num, T* param_out,
                            const MT* master_param, MT* master_param_out) {
  MT lr = static_cast<MT>(learning_rate[0]);
30
  CUDA_KERNEL_LOOP(i, num) {
31 32 33 34 35 36 37
    MT p_data = master_param ? master_param[i] : static_cast<MT>(param[i]);
    MT g_data = static_cast<MT>(grad[i]);
    p_data = p_data - lr * g_data;
    param_out[i] = static_cast<T>(p_data);
    if (master_param_out) {
      master_param_out[i] = p_data;
    }
C
chengduoZH 已提交
38 39 40
  }
}

C
chengduo 已提交
41
template <typename T>
Q
qijun 已提交
42 43 44
__global__ void SparseSGDFunctorKernel(const T* selected_rows,
                                       const int64_t* rows,
                                       const T* learning_rate, T* tensor_out,
C
chengduo 已提交
45 46 47 48 49 50 51 52 53
                                       int64_t row_numel, int64_t limit) {
  for (int64_t i = blockIdx.x; i < limit; i += gridDim.x) {
    const T* selected_rows_ptr = selected_rows + i * row_numel;
    T* tensor_out_ptr = tensor_out + rows[i] * row_numel;
    for (int64_t index = threadIdx.x; index < row_numel; index += blockDim.x) {
      // Since index in rows of SelectedRows can be duplicate, we have to use
      // Atomic Operation to avoid concurrent write error.
      paddle::platform::CudaAtomicAdd(
          tensor_out_ptr + index,
54
          -static_cast<T>(1.0) * learning_rate[0] * selected_rows_ptr[index]);
C
chengduo 已提交
55
    }
Q
qijun 已提交
56 57 58 59 60
  }
}
}  // namespace

template <typename T>
61 62
class SGDOpKernel<platform::CUDADeviceContext, T>
    : public framework::OpKernel<T> {
C
chengduoZH 已提交
63 64
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
C
chengduo 已提交
65
    const auto* param_var = ctx.InputVar("Param");
C
Chengmo 已提交
66 67 68 69 70 71
    PADDLE_ENFORCE_EQ(param_var->IsType<framework::LoDTensor>(), true,
                      platform::errors::InvalidArgument(
                          "The Var(%s)'s type should be LoDTensor, "
                          "but the received is %s",
                          ctx.InputNames("Param").front(),
                          paddle::framework::ToTypeName(param_var->Type())));
72 73
    using paddle::framework::Tensor;
    using MPDType = typename details::MPTypeTrait<T>::Type;
C
chengduo 已提交
74

C
chengduoZH 已提交
75 76 77 78 79
    auto* param = ctx.Input<framework::Tensor>("Param");
    auto* param_out = ctx.Output<framework::Tensor>("ParamOut");
    auto* learning_rate = ctx.Input<framework::Tensor>("LearningRate");

    auto* grad_var = ctx.InputVar("Grad");
80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101

    const bool multi_precision = ctx.Attr<bool>("multi_precision");
    const Tensor* master_param = nullptr;
    Tensor* master_param_out = nullptr;
    if (multi_precision) {
      bool has_master =
          ctx.HasInput("MasterParam") && ctx.HasOutput("MasterParamOut");
      PADDLE_ENFORCE_EQ(has_master, true,
                        platform::errors::InvalidArgument(
                            "The Input(MasterParam) and Output(MasterParamOut) "
                            "should not be null when "
                            "the attr `multi_precision` is true"));
      master_param = ctx.Input<framework::Tensor>("MasterParam");
      master_param_out = ctx.Output<framework::Tensor>("MasterParamOut");
    }
    const MPDType* master_in_data =
        multi_precision ? master_param->data<MPDType>() : nullptr;
    MPDType* master_out_data =
        multi_precision
            ? master_param_out->mutable_data<MPDType>(ctx.GetPlace())
            : nullptr;

C
chengduoZH 已提交
102 103 104 105 106 107 108
    // Actually, all tensors are LoDTensor except SelectedRows.
    if (grad_var->IsType<framework::LoDTensor>()) {
      auto* grad = ctx.Input<framework::Tensor>("Grad");

      int block = 512;
      int grid = (param->numel() + block - 1) / block;

109 110 111 112 113
      SGDKernelMT<
          T, MPDType><<<grid, block, 0, ctx.cuda_device_context().stream()>>>(
          param->data<T>(), grad->data<T>(), learning_rate->data<T>(),
          param->numel(), param_out->mutable_data<T>(ctx.GetPlace()),
          master_in_data, master_out_data);
C
chengduoZH 已提交
114

115
    } else if (grad_var->IsType<pten::SelectedRows>()) {
C
chengduoZH 已提交
116 117 118
      // TODO(qijun): In Sparse SGD operator, in-place update is enforced.
      // This manual optimization brings difficulty to track data dependency.
      // It's better to find a more elegant solution.
C
Chengmo 已提交
119 120 121 122 123
      PADDLE_ENFORCE_EQ(
          param, param_out,
          platform::errors::InvalidArgument(
              "The input tensor Param of SgdOp should be equal with ParamOut "
              "if variable's type is SelectedRows."));
124
      auto* grad = ctx.Input<pten::SelectedRows>("Grad");
C
chengduoZH 已提交
125 126 127

      auto in_height = grad->height();
      auto out_dims = param_out->dims();
C
Chengmo 已提交
128 129 130 131 132 133
      PADDLE_ENFORCE_EQ(in_height, out_dims[0],
                        platform::errors::InvalidArgument(
                            "The input tensor Grad's height of SgdOp should be "
                            "equal with ParamOut's dims. But received Grad's "
                            "height [%s] and ParamOut's dims [%s]",
                            in_height, out_dims[0]));
C
chengduoZH 已提交
134 135

      auto& in_value = grad->value();
Y
Yu Yang 已提交
136
      auto& in_rows = grad->rows();
C
chengduoZH 已提交
137 138

      int64_t in_row_numel = in_value.numel() / in_rows.size();
C
Chengmo 已提交
139 140 141 142
      PADDLE_ENFORCE_EQ(in_row_numel, param_out->numel() / in_height,
                        platform::errors::InvalidArgument(
                            "The in_row_numel of SgdOp should be equal with "
                            "param_out's numel / in_height."));
C
chengduoZH 已提交
143 144 145 146

      auto* in_data = in_value.data<T>();
      auto* out_data = param_out->data<T>();

C
chengduo 已提交
147 148 149 150 151 152 153
      const int kThreadsPerBlock = 256;
      int thread_x = kThreadsPerBlock;
      int max_threads = ctx.cuda_device_context().GetMaxPhysicalThreadCount();
      int max_blocks = std::max(max_threads / kThreadsPerBlock, 1);

      SparseSGDFunctorKernel<<<max_blocks, thread_x, 0,
                               ctx.cuda_device_context().stream()>>>(
Y
Yu Yang 已提交
154
          in_data, in_rows.CUDAData(ctx.GetPlace()), learning_rate->data<T>(),
C
chengduo 已提交
155
          out_data, in_row_numel, in_rows.size());
C
chengduoZH 已提交
156 157

    } else {
C
Chengmo 已提交
158 159 160 161 162 163
      PADDLE_ENFORCE_EQ(false, true,
                        platform::errors::PermissionDenied(
                            "Unsupported Variable Type of Grad "
                            "in SgdOp. Excepted LodTensor or "
                            "SelectedRows, But received [%s]",
                            paddle::framework::ToTypeName(grad_var->Type())));
C
chengduoZH 已提交
164
    }
Q
qijun 已提交
165 166 167 168
  }
};
}  // namespace operators
}  // namespace paddle
Q
Qiao Longfei 已提交
169

D
dongzhihong 已提交
170
namespace ops = paddle::operators;
171
namespace plat = paddle::platform;
172 173 174 175
REGISTER_OP_CUDA_KERNEL(
    sgd, ops::SGDOpKernel<paddle::platform::CUDADeviceContext, float>,
    ops::SGDOpKernel<paddle::platform::CUDADeviceContext, double>,
    ops::SGDOpKernel<paddle::platform::CUDADeviceContext, plat::float16>);